1use ai_onboarding::YoungAccountBanner;
2use anthropic::AnthropicModelMode;
3use anyhow::{Context as _, Result, anyhow};
4use chrono::{DateTime, Utc};
5use client::{Client, ModelRequestUsage, UserStore, zed_urls};
6use cloud_llm_client::{
7 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
8 CompletionEvent, CompletionRequestStatus, CountTokensBody, CountTokensResponse,
9 EXPIRED_LLM_TOKEN_HEADER_NAME, ListModelsResponse, MODEL_REQUESTS_RESOURCE_HEADER_VALUE, Plan,
10 PlanV1, PlanV2, SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME,
11 SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME, TOOL_USE_LIMIT_REACHED_HEADER_NAME,
12 ZED_VERSION_HEADER_NAME,
13};
14use futures::{
15 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
16};
17use google_ai::GoogleModelMode;
18use gpui::{
19 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
20};
21use http_client::http::{HeaderMap, HeaderValue};
22use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
23use language_model::{
24 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
25 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
26 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
27 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
28 LanguageModelToolSchemaFormat, LlmApiToken, ModelRequestLimitReachedError,
29 PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
30};
31use release_channel::AppVersion;
32use schemars::JsonSchema;
33use serde::{Deserialize, Serialize, de::DeserializeOwned};
34use settings::SettingsStore;
35use smol::io::{AsyncReadExt, BufReader};
36use std::pin::Pin;
37use std::str::FromStr as _;
38use std::sync::Arc;
39use std::time::Duration;
40use thiserror::Error;
41use ui::{TintColor, prelude::*};
42use util::{ResultExt as _, maybe};
43
44use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
45use crate::provider::google::{GoogleEventMapper, into_google};
46use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
47
48const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
49const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
50
51#[derive(Default, Clone, Debug, PartialEq)]
52pub struct ZedDotDevSettings {
53 pub available_models: Vec<AvailableModel>,
54}
55
56#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
57#[serde(rename_all = "lowercase")]
58pub enum AvailableProvider {
59 Anthropic,
60 OpenAi,
61 Google,
62}
63
64#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
65pub struct AvailableModel {
66 /// The provider of the language model.
67 pub provider: AvailableProvider,
68 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
69 pub name: String,
70 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
71 pub display_name: Option<String>,
72 /// The size of the context window, indicating the maximum number of tokens the model can process.
73 pub max_tokens: usize,
74 /// The maximum number of output tokens allowed by the model.
75 pub max_output_tokens: Option<u64>,
76 /// The maximum number of completion tokens allowed by the model (o1-* only)
77 pub max_completion_tokens: Option<u64>,
78 /// Override this model with a different Anthropic model for tool calls.
79 pub tool_override: Option<String>,
80 /// Indicates whether this custom model supports caching.
81 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
82 /// The default temperature to use for this model.
83 pub default_temperature: Option<f32>,
84 /// Any extra beta headers to provide when using the model.
85 #[serde(default)]
86 pub extra_beta_headers: Vec<String>,
87 /// The model's mode (e.g. thinking)
88 pub mode: Option<ModelMode>,
89}
90
91#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
92#[serde(tag = "type", rename_all = "lowercase")]
93pub enum ModelMode {
94 #[default]
95 Default,
96 Thinking {
97 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
98 budget_tokens: Option<u32>,
99 },
100}
101
102impl From<ModelMode> for AnthropicModelMode {
103 fn from(value: ModelMode) -> Self {
104 match value {
105 ModelMode::Default => AnthropicModelMode::Default,
106 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
107 }
108 }
109}
110
111pub struct CloudLanguageModelProvider {
112 client: Arc<Client>,
113 state: gpui::Entity<State>,
114 _maintain_client_status: Task<()>,
115}
116
117pub struct State {
118 client: Arc<Client>,
119 llm_api_token: LlmApiToken,
120 user_store: Entity<UserStore>,
121 status: client::Status,
122 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
123 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
124 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
125 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
126 _fetch_models_task: Task<()>,
127 _settings_subscription: Subscription,
128 _llm_token_subscription: Subscription,
129}
130
131impl State {
132 fn new(
133 client: Arc<Client>,
134 user_store: Entity<UserStore>,
135 status: client::Status,
136 cx: &mut Context<Self>,
137 ) -> Self {
138 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
139 let mut current_user = user_store.read(cx).watch_current_user();
140 Self {
141 client: client.clone(),
142 llm_api_token: LlmApiToken::default(),
143 user_store,
144 status,
145 models: Vec::new(),
146 default_model: None,
147 default_fast_model: None,
148 recommended_models: Vec::new(),
149 _fetch_models_task: cx.spawn(async move |this, cx| {
150 maybe!(async move {
151 let (client, llm_api_token) = this
152 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
153
154 while current_user.borrow().is_none() {
155 current_user.next().await;
156 }
157
158 let response =
159 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
160 this.update(cx, |this, cx| this.update_models(response, cx))?;
161 anyhow::Ok(())
162 })
163 .await
164 .context("failed to fetch Zed models")
165 .log_err();
166 }),
167 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
168 cx.notify();
169 }),
170 _llm_token_subscription: cx.subscribe(
171 &refresh_llm_token_listener,
172 move |this, _listener, _event, cx| {
173 let client = this.client.clone();
174 let llm_api_token = this.llm_api_token.clone();
175 cx.spawn(async move |this, cx| {
176 llm_api_token.refresh(&client).await?;
177 let response = Self::fetch_models(client, llm_api_token).await?;
178 this.update(cx, |this, cx| {
179 this.update_models(response, cx);
180 })
181 })
182 .detach_and_log_err(cx);
183 },
184 ),
185 }
186 }
187
188 fn is_signed_out(&self, cx: &App) -> bool {
189 self.user_store.read(cx).current_user().is_none()
190 }
191
192 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
193 let client = self.client.clone();
194 cx.spawn(async move |state, cx| {
195 client.sign_in_with_optional_connect(true, cx).await?;
196 state.update(cx, |_, cx| cx.notify())
197 })
198 }
199 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
200 let mut models = Vec::new();
201
202 for model in response.models {
203 models.push(Arc::new(model.clone()));
204
205 // Right now we represent thinking variants of models as separate models on the client,
206 // so we need to insert variants for any model that supports thinking.
207 if model.supports_thinking {
208 models.push(Arc::new(cloud_llm_client::LanguageModel {
209 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
210 display_name: format!("{} Thinking", model.display_name),
211 ..model
212 }));
213 }
214 }
215
216 self.default_model = models
217 .iter()
218 .find(|model| model.id == response.default_model)
219 .cloned();
220 self.default_fast_model = models
221 .iter()
222 .find(|model| model.id == response.default_fast_model)
223 .cloned();
224 self.recommended_models = response
225 .recommended_models
226 .iter()
227 .filter_map(|id| models.iter().find(|model| &model.id == id))
228 .cloned()
229 .collect();
230 self.models = models;
231 cx.notify();
232 }
233
234 async fn fetch_models(
235 client: Arc<Client>,
236 llm_api_token: LlmApiToken,
237 ) -> Result<ListModelsResponse> {
238 let http_client = &client.http_client();
239 let token = llm_api_token.acquire(&client).await?;
240
241 let request = http_client::Request::builder()
242 .method(Method::GET)
243 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
244 .header("Authorization", format!("Bearer {token}"))
245 .body(AsyncBody::empty())?;
246 let mut response = http_client
247 .send(request)
248 .await
249 .context("failed to send list models request")?;
250
251 if response.status().is_success() {
252 let mut body = String::new();
253 response.body_mut().read_to_string(&mut body).await?;
254 Ok(serde_json::from_str(&body)?)
255 } else {
256 let mut body = String::new();
257 response.body_mut().read_to_string(&mut body).await?;
258 anyhow::bail!(
259 "error listing models.\nStatus: {:?}\nBody: {body}",
260 response.status(),
261 );
262 }
263 }
264}
265
266impl CloudLanguageModelProvider {
267 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
268 let mut status_rx = client.status();
269 let status = *status_rx.borrow();
270
271 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
272
273 let state_ref = state.downgrade();
274 let maintain_client_status = cx.spawn(async move |cx| {
275 while let Some(status) = status_rx.next().await {
276 if let Some(this) = state_ref.upgrade() {
277 _ = this.update(cx, |this, cx| {
278 if this.status != status {
279 this.status = status;
280 cx.notify();
281 }
282 });
283 } else {
284 break;
285 }
286 }
287 });
288
289 Self {
290 client,
291 state,
292 _maintain_client_status: maintain_client_status,
293 }
294 }
295
296 fn create_language_model(
297 &self,
298 model: Arc<cloud_llm_client::LanguageModel>,
299 llm_api_token: LlmApiToken,
300 ) -> Arc<dyn LanguageModel> {
301 Arc::new(CloudLanguageModel {
302 id: LanguageModelId(SharedString::from(model.id.0.clone())),
303 model,
304 llm_api_token,
305 client: self.client.clone(),
306 request_limiter: RateLimiter::new(4),
307 })
308 }
309}
310
311impl LanguageModelProviderState for CloudLanguageModelProvider {
312 type ObservableEntity = State;
313
314 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
315 Some(self.state.clone())
316 }
317}
318
319impl LanguageModelProvider for CloudLanguageModelProvider {
320 fn id(&self) -> LanguageModelProviderId {
321 PROVIDER_ID
322 }
323
324 fn name(&self) -> LanguageModelProviderName {
325 PROVIDER_NAME
326 }
327
328 fn icon(&self) -> IconName {
329 IconName::AiZed
330 }
331
332 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
333 let default_model = self.state.read(cx).default_model.clone()?;
334 let llm_api_token = self.state.read(cx).llm_api_token.clone();
335 Some(self.create_language_model(default_model, llm_api_token))
336 }
337
338 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
339 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
340 let llm_api_token = self.state.read(cx).llm_api_token.clone();
341 Some(self.create_language_model(default_fast_model, llm_api_token))
342 }
343
344 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
345 let llm_api_token = self.state.read(cx).llm_api_token.clone();
346 self.state
347 .read(cx)
348 .recommended_models
349 .iter()
350 .cloned()
351 .map(|model| self.create_language_model(model, llm_api_token.clone()))
352 .collect()
353 }
354
355 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
356 let llm_api_token = self.state.read(cx).llm_api_token.clone();
357 self.state
358 .read(cx)
359 .models
360 .iter()
361 .cloned()
362 .map(|model| self.create_language_model(model, llm_api_token.clone()))
363 .collect()
364 }
365
366 fn is_authenticated(&self, cx: &App) -> bool {
367 let state = self.state.read(cx);
368 !state.is_signed_out(cx)
369 }
370
371 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
372 Task::ready(Ok(()))
373 }
374
375 fn configuration_view(
376 &self,
377 _target_agent: language_model::ConfigurationViewTargetAgent,
378 _: &mut Window,
379 cx: &mut App,
380 ) -> AnyView {
381 cx.new(|_| ConfigurationView::new(self.state.clone()))
382 .into()
383 }
384
385 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
386 Task::ready(Ok(()))
387 }
388}
389
390pub struct CloudLanguageModel {
391 id: LanguageModelId,
392 model: Arc<cloud_llm_client::LanguageModel>,
393 llm_api_token: LlmApiToken,
394 client: Arc<Client>,
395 request_limiter: RateLimiter,
396}
397
398struct PerformLlmCompletionResponse {
399 response: Response<AsyncBody>,
400 usage: Option<ModelRequestUsage>,
401 tool_use_limit_reached: bool,
402 includes_status_messages: bool,
403}
404
405impl CloudLanguageModel {
406 async fn perform_llm_completion(
407 client: Arc<Client>,
408 llm_api_token: LlmApiToken,
409 app_version: Option<SemanticVersion>,
410 body: CompletionBody,
411 ) -> Result<PerformLlmCompletionResponse> {
412 let http_client = &client.http_client();
413
414 let mut token = llm_api_token.acquire(&client).await?;
415 let mut refreshed_token = false;
416
417 loop {
418 let request_builder = http_client::Request::builder()
419 .method(Method::POST)
420 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
421 let request_builder = if let Some(app_version) = app_version {
422 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
423 } else {
424 request_builder
425 };
426
427 let request = request_builder
428 .header("Content-Type", "application/json")
429 .header("Authorization", format!("Bearer {token}"))
430 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
431 .body(serde_json::to_string(&body)?.into())?;
432 let mut response = http_client.send(request).await?;
433 let status = response.status();
434 if status.is_success() {
435 let includes_status_messages = response
436 .headers()
437 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
438 .is_some();
439
440 let tool_use_limit_reached = response
441 .headers()
442 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
443 .is_some();
444
445 let usage = if includes_status_messages {
446 None
447 } else {
448 ModelRequestUsage::from_headers(response.headers()).ok()
449 };
450
451 return Ok(PerformLlmCompletionResponse {
452 response,
453 usage,
454 includes_status_messages,
455 tool_use_limit_reached,
456 });
457 }
458
459 if !refreshed_token
460 && response
461 .headers()
462 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
463 .is_some()
464 {
465 token = llm_api_token.refresh(&client).await?;
466 refreshed_token = true;
467 continue;
468 }
469
470 if status == StatusCode::FORBIDDEN
471 && response
472 .headers()
473 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
474 .is_some()
475 {
476 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
477 .headers()
478 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
479 .and_then(|resource| resource.to_str().ok())
480 && let Some(plan) = response
481 .headers()
482 .get(CURRENT_PLAN_HEADER_NAME)
483 .and_then(|plan| plan.to_str().ok())
484 .and_then(|plan| cloud_llm_client::PlanV1::from_str(plan).ok())
485 .map(Plan::V1)
486 {
487 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
488 }
489 } else if status == StatusCode::PAYMENT_REQUIRED {
490 return Err(anyhow!(PaymentRequiredError));
491 }
492
493 let mut body = String::new();
494 let headers = response.headers().clone();
495 response.body_mut().read_to_string(&mut body).await?;
496 return Err(anyhow!(ApiError {
497 status,
498 body,
499 headers
500 }));
501 }
502 }
503}
504
505#[derive(Debug, Error)]
506#[error("cloud language model request failed with status {status}: {body}")]
507struct ApiError {
508 status: StatusCode,
509 body: String,
510 headers: HeaderMap<HeaderValue>,
511}
512
513/// Represents error responses from Zed's cloud API.
514///
515/// Example JSON for an upstream HTTP error:
516/// ```json
517/// {
518/// "code": "upstream_http_error",
519/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
520/// "upstream_status": 503
521/// }
522/// ```
523#[derive(Debug, serde::Deserialize)]
524struct CloudApiError {
525 code: String,
526 message: String,
527 #[serde(default)]
528 #[serde(deserialize_with = "deserialize_optional_status_code")]
529 upstream_status: Option<StatusCode>,
530 #[serde(default)]
531 retry_after: Option<f64>,
532}
533
534fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
535where
536 D: serde::Deserializer<'de>,
537{
538 let opt: Option<u16> = Option::deserialize(deserializer)?;
539 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
540}
541
542impl From<ApiError> for LanguageModelCompletionError {
543 fn from(error: ApiError) -> Self {
544 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
545 if cloud_error.code.starts_with("upstream_http_") {
546 let status = if let Some(status) = cloud_error.upstream_status {
547 status
548 } else if cloud_error.code.ends_with("_error") {
549 error.status
550 } else {
551 // If there's a status code in the code string (e.g. "upstream_http_429")
552 // then use that; otherwise, see if the JSON contains a status code.
553 cloud_error
554 .code
555 .strip_prefix("upstream_http_")
556 .and_then(|code_str| code_str.parse::<u16>().ok())
557 .and_then(|code| StatusCode::from_u16(code).ok())
558 .unwrap_or(error.status)
559 };
560
561 return LanguageModelCompletionError::UpstreamProviderError {
562 message: cloud_error.message,
563 status,
564 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
565 };
566 }
567
568 return LanguageModelCompletionError::from_http_status(
569 PROVIDER_NAME,
570 error.status,
571 cloud_error.message,
572 None,
573 );
574 }
575
576 let retry_after = None;
577 LanguageModelCompletionError::from_http_status(
578 PROVIDER_NAME,
579 error.status,
580 error.body,
581 retry_after,
582 )
583 }
584}
585
586impl LanguageModel for CloudLanguageModel {
587 fn id(&self) -> LanguageModelId {
588 self.id.clone()
589 }
590
591 fn name(&self) -> LanguageModelName {
592 LanguageModelName::from(self.model.display_name.clone())
593 }
594
595 fn provider_id(&self) -> LanguageModelProviderId {
596 PROVIDER_ID
597 }
598
599 fn provider_name(&self) -> LanguageModelProviderName {
600 PROVIDER_NAME
601 }
602
603 fn upstream_provider_id(&self) -> LanguageModelProviderId {
604 use cloud_llm_client::LanguageModelProvider::*;
605 match self.model.provider {
606 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
607 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
608 Google => language_model::GOOGLE_PROVIDER_ID,
609 }
610 }
611
612 fn upstream_provider_name(&self) -> LanguageModelProviderName {
613 use cloud_llm_client::LanguageModelProvider::*;
614 match self.model.provider {
615 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
616 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
617 Google => language_model::GOOGLE_PROVIDER_NAME,
618 }
619 }
620
621 fn supports_tools(&self) -> bool {
622 self.model.supports_tools
623 }
624
625 fn supports_images(&self) -> bool {
626 self.model.supports_images
627 }
628
629 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
630 match choice {
631 LanguageModelToolChoice::Auto
632 | LanguageModelToolChoice::Any
633 | LanguageModelToolChoice::None => true,
634 }
635 }
636
637 fn supports_burn_mode(&self) -> bool {
638 self.model.supports_max_mode
639 }
640
641 fn telemetry_id(&self) -> String {
642 format!("zed.dev/{}", self.model.id)
643 }
644
645 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
646 match self.model.provider {
647 cloud_llm_client::LanguageModelProvider::Anthropic
648 | cloud_llm_client::LanguageModelProvider::OpenAi => {
649 LanguageModelToolSchemaFormat::JsonSchema
650 }
651 cloud_llm_client::LanguageModelProvider::Google => {
652 LanguageModelToolSchemaFormat::JsonSchemaSubset
653 }
654 }
655 }
656
657 fn max_token_count(&self) -> u64 {
658 self.model.max_token_count as u64
659 }
660
661 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
662 self.model
663 .max_token_count_in_max_mode
664 .filter(|_| self.model.supports_max_mode)
665 .map(|max_token_count| max_token_count as u64)
666 }
667
668 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
669 match &self.model.provider {
670 cloud_llm_client::LanguageModelProvider::Anthropic => {
671 Some(LanguageModelCacheConfiguration {
672 min_total_token: 2_048,
673 should_speculate: true,
674 max_cache_anchors: 4,
675 })
676 }
677 cloud_llm_client::LanguageModelProvider::OpenAi
678 | cloud_llm_client::LanguageModelProvider::Google => None,
679 }
680 }
681
682 fn count_tokens(
683 &self,
684 request: LanguageModelRequest,
685 cx: &App,
686 ) -> BoxFuture<'static, Result<u64>> {
687 match self.model.provider {
688 cloud_llm_client::LanguageModelProvider::Anthropic => {
689 count_anthropic_tokens(request, cx)
690 }
691 cloud_llm_client::LanguageModelProvider::OpenAi => {
692 let model = match open_ai::Model::from_id(&self.model.id.0) {
693 Ok(model) => model,
694 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
695 };
696 count_open_ai_tokens(request, model, cx)
697 }
698 cloud_llm_client::LanguageModelProvider::Google => {
699 let client = self.client.clone();
700 let llm_api_token = self.llm_api_token.clone();
701 let model_id = self.model.id.to_string();
702 let generate_content_request =
703 into_google(request, model_id.clone(), GoogleModelMode::Default);
704 async move {
705 let http_client = &client.http_client();
706 let token = llm_api_token.acquire(&client).await?;
707
708 let request_body = CountTokensBody {
709 provider: cloud_llm_client::LanguageModelProvider::Google,
710 model: model_id,
711 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
712 generate_content_request,
713 })?,
714 };
715 let request = http_client::Request::builder()
716 .method(Method::POST)
717 .uri(
718 http_client
719 .build_zed_llm_url("/count_tokens", &[])?
720 .as_ref(),
721 )
722 .header("Content-Type", "application/json")
723 .header("Authorization", format!("Bearer {token}"))
724 .body(serde_json::to_string(&request_body)?.into())?;
725 let mut response = http_client.send(request).await?;
726 let status = response.status();
727 let headers = response.headers().clone();
728 let mut response_body = String::new();
729 response
730 .body_mut()
731 .read_to_string(&mut response_body)
732 .await?;
733
734 if status.is_success() {
735 let response_body: CountTokensResponse =
736 serde_json::from_str(&response_body)?;
737
738 Ok(response_body.tokens as u64)
739 } else {
740 Err(anyhow!(ApiError {
741 status,
742 body: response_body,
743 headers
744 }))
745 }
746 }
747 .boxed()
748 }
749 }
750 }
751
752 fn stream_completion(
753 &self,
754 request: LanguageModelRequest,
755 cx: &AsyncApp,
756 ) -> BoxFuture<
757 'static,
758 Result<
759 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
760 LanguageModelCompletionError,
761 >,
762 > {
763 let thread_id = request.thread_id.clone();
764 let prompt_id = request.prompt_id.clone();
765 let intent = request.intent;
766 let mode = request.mode;
767 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
768 let thinking_allowed = request.thinking_allowed;
769 match self.model.provider {
770 cloud_llm_client::LanguageModelProvider::Anthropic => {
771 let request = into_anthropic(
772 request,
773 self.model.id.to_string(),
774 1.0,
775 self.model.max_output_tokens as u64,
776 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
777 AnthropicModelMode::Thinking {
778 budget_tokens: Some(4_096),
779 }
780 } else {
781 AnthropicModelMode::Default
782 },
783 );
784 let client = self.client.clone();
785 let llm_api_token = self.llm_api_token.clone();
786 let future = self.request_limiter.stream(async move {
787 let PerformLlmCompletionResponse {
788 response,
789 usage,
790 includes_status_messages,
791 tool_use_limit_reached,
792 } = Self::perform_llm_completion(
793 client.clone(),
794 llm_api_token,
795 app_version,
796 CompletionBody {
797 thread_id,
798 prompt_id,
799 intent,
800 mode,
801 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
802 model: request.model.clone(),
803 provider_request: serde_json::to_value(&request)
804 .map_err(|e| anyhow!(e))?,
805 },
806 )
807 .await
808 .map_err(|err| match err.downcast::<ApiError>() {
809 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
810 Err(err) => anyhow!(err),
811 })?;
812
813 let mut mapper = AnthropicEventMapper::new();
814 Ok(map_cloud_completion_events(
815 Box::pin(
816 response_lines(response, includes_status_messages)
817 .chain(usage_updated_event(usage))
818 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
819 ),
820 move |event| mapper.map_event(event),
821 ))
822 });
823 async move { Ok(future.await?.boxed()) }.boxed()
824 }
825 cloud_llm_client::LanguageModelProvider::OpenAi => {
826 let client = self.client.clone();
827 let model = match open_ai::Model::from_id(&self.model.id.0) {
828 Ok(model) => model,
829 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
830 };
831 let request = into_open_ai(
832 request,
833 model.id(),
834 model.supports_parallel_tool_calls(),
835 model.supports_prompt_cache_key(),
836 None,
837 None,
838 );
839 let llm_api_token = self.llm_api_token.clone();
840 let future = self.request_limiter.stream(async move {
841 let PerformLlmCompletionResponse {
842 response,
843 usage,
844 includes_status_messages,
845 tool_use_limit_reached,
846 } = Self::perform_llm_completion(
847 client.clone(),
848 llm_api_token,
849 app_version,
850 CompletionBody {
851 thread_id,
852 prompt_id,
853 intent,
854 mode,
855 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
856 model: request.model.clone(),
857 provider_request: serde_json::to_value(&request)
858 .map_err(|e| anyhow!(e))?,
859 },
860 )
861 .await?;
862
863 let mut mapper = OpenAiEventMapper::new();
864 Ok(map_cloud_completion_events(
865 Box::pin(
866 response_lines(response, includes_status_messages)
867 .chain(usage_updated_event(usage))
868 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
869 ),
870 move |event| mapper.map_event(event),
871 ))
872 });
873 async move { Ok(future.await?.boxed()) }.boxed()
874 }
875 cloud_llm_client::LanguageModelProvider::Google => {
876 let client = self.client.clone();
877 let request =
878 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
879 let llm_api_token = self.llm_api_token.clone();
880 let future = self.request_limiter.stream(async move {
881 let PerformLlmCompletionResponse {
882 response,
883 usage,
884 includes_status_messages,
885 tool_use_limit_reached,
886 } = Self::perform_llm_completion(
887 client.clone(),
888 llm_api_token,
889 app_version,
890 CompletionBody {
891 thread_id,
892 prompt_id,
893 intent,
894 mode,
895 provider: cloud_llm_client::LanguageModelProvider::Google,
896 model: request.model.model_id.clone(),
897 provider_request: serde_json::to_value(&request)
898 .map_err(|e| anyhow!(e))?,
899 },
900 )
901 .await?;
902
903 let mut mapper = GoogleEventMapper::new();
904 Ok(map_cloud_completion_events(
905 Box::pin(
906 response_lines(response, includes_status_messages)
907 .chain(usage_updated_event(usage))
908 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
909 ),
910 move |event| mapper.map_event(event),
911 ))
912 });
913 async move { Ok(future.await?.boxed()) }.boxed()
914 }
915 }
916 }
917}
918
919fn map_cloud_completion_events<T, F>(
920 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
921 mut map_callback: F,
922) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
923where
924 T: DeserializeOwned + 'static,
925 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
926 + Send
927 + 'static,
928{
929 stream
930 .flat_map(move |event| {
931 futures::stream::iter(match event {
932 Err(error) => {
933 vec![Err(LanguageModelCompletionError::from(error))]
934 }
935 Ok(CompletionEvent::Status(event)) => {
936 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
937 }
938 Ok(CompletionEvent::Event(event)) => map_callback(event),
939 })
940 })
941 .boxed()
942}
943
944fn usage_updated_event<T>(
945 usage: Option<ModelRequestUsage>,
946) -> impl Stream<Item = Result<CompletionEvent<T>>> {
947 futures::stream::iter(usage.map(|usage| {
948 Ok(CompletionEvent::Status(
949 CompletionRequestStatus::UsageUpdated {
950 amount: usage.amount as usize,
951 limit: usage.limit,
952 },
953 ))
954 }))
955}
956
957fn tool_use_limit_reached_event<T>(
958 tool_use_limit_reached: bool,
959) -> impl Stream<Item = Result<CompletionEvent<T>>> {
960 futures::stream::iter(tool_use_limit_reached.then(|| {
961 Ok(CompletionEvent::Status(
962 CompletionRequestStatus::ToolUseLimitReached,
963 ))
964 }))
965}
966
967fn response_lines<T: DeserializeOwned>(
968 response: Response<AsyncBody>,
969 includes_status_messages: bool,
970) -> impl Stream<Item = Result<CompletionEvent<T>>> {
971 futures::stream::try_unfold(
972 (String::new(), BufReader::new(response.into_body())),
973 move |(mut line, mut body)| async move {
974 match body.read_line(&mut line).await {
975 Ok(0) => Ok(None),
976 Ok(_) => {
977 let event = if includes_status_messages {
978 serde_json::from_str::<CompletionEvent<T>>(&line)?
979 } else {
980 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
981 };
982
983 line.clear();
984 Ok(Some((event, (line, body))))
985 }
986 Err(e) => Err(e.into()),
987 }
988 },
989 )
990}
991
992#[derive(IntoElement, RegisterComponent)]
993struct ZedAiConfiguration {
994 is_connected: bool,
995 plan: Option<Plan>,
996 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
997 eligible_for_trial: bool,
998 account_too_young: bool,
999 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1000}
1001
1002impl RenderOnce for ZedAiConfiguration {
1003 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1004 let young_account_banner = YoungAccountBanner;
1005
1006 let is_pro = self.plan.is_some_and(|plan| {
1007 matches!(plan, Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro))
1008 });
1009 let subscription_text = match (self.plan, self.subscription_period) {
1010 (Some(Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1011 "You have access to Zed's hosted models through your Pro subscription."
1012 }
1013 (Some(Plan::V1(PlanV1::ZedProTrial) | Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1014 "You have access to Zed's hosted models through your Pro trial."
1015 }
1016 (Some(Plan::V1(PlanV1::ZedFree) | Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1017 "You have basic access to Zed's hosted models through the Free plan."
1018 }
1019 _ => {
1020 if self.eligible_for_trial {
1021 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1022 } else {
1023 "Subscribe for access to Zed's hosted models."
1024 }
1025 }
1026 };
1027
1028 let manage_subscription_buttons = if is_pro {
1029 Button::new("manage_settings", "Manage Subscription")
1030 .full_width()
1031 .style(ButtonStyle::Tinted(TintColor::Accent))
1032 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1033 .into_any_element()
1034 } else if self.plan.is_none() || self.eligible_for_trial {
1035 Button::new("start_trial", "Start 14-day Free Pro Trial")
1036 .full_width()
1037 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1038 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1039 .into_any_element()
1040 } else {
1041 Button::new("upgrade", "Upgrade to Pro")
1042 .full_width()
1043 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1044 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1045 .into_any_element()
1046 };
1047
1048 if !self.is_connected {
1049 return v_flex()
1050 .gap_2()
1051 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1052 .child(
1053 Button::new("sign_in", "Sign In to use Zed AI")
1054 .icon_color(Color::Muted)
1055 .icon(IconName::Github)
1056 .icon_size(IconSize::Small)
1057 .icon_position(IconPosition::Start)
1058 .full_width()
1059 .on_click({
1060 let callback = self.sign_in_callback.clone();
1061 move |_, window, cx| (callback)(window, cx)
1062 }),
1063 );
1064 }
1065
1066 v_flex().gap_2().w_full().map(|this| {
1067 if self.account_too_young {
1068 this.child(young_account_banner).child(
1069 Button::new("upgrade", "Upgrade to Pro")
1070 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1071 .full_width()
1072 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1073 )
1074 } else {
1075 this.text_sm()
1076 .child(subscription_text)
1077 .child(manage_subscription_buttons)
1078 }
1079 })
1080 }
1081}
1082
1083struct ConfigurationView {
1084 state: Entity<State>,
1085 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1086}
1087
1088impl ConfigurationView {
1089 fn new(state: Entity<State>) -> Self {
1090 let sign_in_callback = Arc::new({
1091 let state = state.clone();
1092 move |_window: &mut Window, cx: &mut App| {
1093 state.update(cx, |state, cx| {
1094 state.authenticate(cx).detach_and_log_err(cx);
1095 });
1096 }
1097 });
1098
1099 Self {
1100 state,
1101 sign_in_callback,
1102 }
1103 }
1104}
1105
1106impl Render for ConfigurationView {
1107 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1108 let state = self.state.read(cx);
1109 let user_store = state.user_store.read(cx);
1110
1111 ZedAiConfiguration {
1112 is_connected: !state.is_signed_out(cx),
1113 plan: user_store.plan(),
1114 subscription_period: user_store.subscription_period(),
1115 eligible_for_trial: user_store.trial_started_at().is_none(),
1116 account_too_young: user_store.account_too_young(),
1117 sign_in_callback: self.sign_in_callback.clone(),
1118 }
1119 }
1120}
1121
1122impl Component for ZedAiConfiguration {
1123 fn name() -> &'static str {
1124 "AI Configuration Content"
1125 }
1126
1127 fn sort_name() -> &'static str {
1128 "AI Configuration Content"
1129 }
1130
1131 fn scope() -> ComponentScope {
1132 ComponentScope::Onboarding
1133 }
1134
1135 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1136 fn configuration(
1137 is_connected: bool,
1138 plan: Option<Plan>,
1139 eligible_for_trial: bool,
1140 account_too_young: bool,
1141 ) -> AnyElement {
1142 ZedAiConfiguration {
1143 is_connected,
1144 plan,
1145 subscription_period: plan
1146 .is_some()
1147 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1148 eligible_for_trial,
1149 account_too_young,
1150 sign_in_callback: Arc::new(|_, _| {}),
1151 }
1152 .into_any_element()
1153 }
1154
1155 Some(
1156 v_flex()
1157 .p_4()
1158 .gap_4()
1159 .children(vec![
1160 single_example("Not connected", configuration(false, None, false, false)),
1161 single_example(
1162 "Accept Terms of Service",
1163 configuration(true, None, true, false),
1164 ),
1165 single_example(
1166 "No Plan - Not eligible for trial",
1167 configuration(true, None, false, false),
1168 ),
1169 single_example(
1170 "No Plan - Eligible for trial",
1171 configuration(true, None, true, false),
1172 ),
1173 single_example(
1174 "Free Plan",
1175 configuration(true, Some(Plan::V1(PlanV1::ZedFree)), true, false),
1176 ),
1177 single_example(
1178 "Zed Pro Trial Plan",
1179 configuration(true, Some(Plan::V1(PlanV1::ZedProTrial)), true, false),
1180 ),
1181 single_example(
1182 "Zed Pro Plan",
1183 configuration(true, Some(Plan::V1(PlanV1::ZedPro)), true, false),
1184 ),
1185 ])
1186 .into_any_element(),
1187 )
1188 }
1189}
1190
1191#[cfg(test)]
1192mod tests {
1193 use super::*;
1194 use http_client::http::{HeaderMap, StatusCode};
1195 use language_model::LanguageModelCompletionError;
1196
1197 #[test]
1198 fn test_api_error_conversion_with_upstream_http_error() {
1199 // upstream_http_error with 503 status should become ServerOverloaded
1200 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout","upstream_status":503}"#;
1201
1202 let api_error = ApiError {
1203 status: StatusCode::INTERNAL_SERVER_ERROR,
1204 body: error_body.to_string(),
1205 headers: HeaderMap::new(),
1206 };
1207
1208 let completion_error: LanguageModelCompletionError = api_error.into();
1209
1210 match completion_error {
1211 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1212 assert_eq!(
1213 message,
1214 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1215 );
1216 }
1217 _ => panic!(
1218 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1219 completion_error
1220 ),
1221 }
1222
1223 // upstream_http_error with 500 status should become ApiInternalServerError
1224 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1225
1226 let api_error = ApiError {
1227 status: StatusCode::INTERNAL_SERVER_ERROR,
1228 body: error_body.to_string(),
1229 headers: HeaderMap::new(),
1230 };
1231
1232 let completion_error: LanguageModelCompletionError = api_error.into();
1233
1234 match completion_error {
1235 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1236 assert_eq!(
1237 message,
1238 "Received an error from the OpenAI API: internal server error"
1239 );
1240 }
1241 _ => panic!(
1242 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1243 completion_error
1244 ),
1245 }
1246
1247 // upstream_http_error with 429 status should become RateLimitExceeded
1248 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1249
1250 let api_error = ApiError {
1251 status: StatusCode::INTERNAL_SERVER_ERROR,
1252 body: error_body.to_string(),
1253 headers: HeaderMap::new(),
1254 };
1255
1256 let completion_error: LanguageModelCompletionError = api_error.into();
1257
1258 match completion_error {
1259 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1260 assert_eq!(
1261 message,
1262 "Received an error from the Google API: rate limit exceeded"
1263 );
1264 }
1265 _ => panic!(
1266 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1267 completion_error
1268 ),
1269 }
1270
1271 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1272 let error_body = "Regular internal server error";
1273
1274 let api_error = ApiError {
1275 status: StatusCode::INTERNAL_SERVER_ERROR,
1276 body: error_body.to_string(),
1277 headers: HeaderMap::new(),
1278 };
1279
1280 let completion_error: LanguageModelCompletionError = api_error.into();
1281
1282 match completion_error {
1283 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1284 assert_eq!(provider, PROVIDER_NAME);
1285 assert_eq!(message, "Regular internal server error");
1286 }
1287 _ => panic!(
1288 "Expected ApiInternalServerError for regular 500, got: {:?}",
1289 completion_error
1290 ),
1291 }
1292
1293 // upstream_http_429 format should be converted to UpstreamProviderError
1294 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1295
1296 let api_error = ApiError {
1297 status: StatusCode::INTERNAL_SERVER_ERROR,
1298 body: error_body.to_string(),
1299 headers: HeaderMap::new(),
1300 };
1301
1302 let completion_error: LanguageModelCompletionError = api_error.into();
1303
1304 match completion_error {
1305 LanguageModelCompletionError::UpstreamProviderError {
1306 message,
1307 status,
1308 retry_after,
1309 } => {
1310 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1311 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1312 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1313 }
1314 _ => panic!(
1315 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1316 completion_error
1317 ),
1318 }
1319
1320 // Invalid JSON in error body should fall back to regular error handling
1321 let error_body = "Not JSON at all";
1322
1323 let api_error = ApiError {
1324 status: StatusCode::INTERNAL_SERVER_ERROR,
1325 body: error_body.to_string(),
1326 headers: HeaderMap::new(),
1327 };
1328
1329 let completion_error: LanguageModelCompletionError = api_error.into();
1330
1331 match completion_error {
1332 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1333 assert_eq!(provider, PROVIDER_NAME);
1334 }
1335 _ => panic!(
1336 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1337 completion_error
1338 ),
1339 }
1340 }
1341}